WTB-Net:基于ShuffleNet V2的轻量级风机叶片表面缺陷识别算法  

WTB-Net:A lightweight wind turbine blade surface defect recognition algorithm based on ShuffleNet V2

作　　者：张睿 文传博[1] ZHANG Rui;WEN Chuanbo(Department of Electrical Engineering,Shanghai Dianji University,Shanghai 201306,China)

机构地区：[1]上海电机学院电气学院,上海201306

出　　处：《复合材料科学与工程》2023年第8期66-71,128,共7页Composites Science and Engineering

基　　金：国家自然科学基金项目(61973209);上海市自然科学基金项目(20ZR1421200);上海市地方高校能力建设项目(22010501100)。

摘　　要：由于现有风机叶片(WTB)的缺陷检测技术尚未得到广泛应用,且传统检测方法的鲁棒性较差,本文提出了一种轻量级风机叶片表面缺陷识别算法WTB-Net。通过无人机拍摄华东某沿海风电场的2569张风机叶片图像,通过筛选分类和数据扩充建立WTB表面缺陷数据集。以ShuffleNet V2核心骨干特征提取网络为基础,引入选择性卷积核注意力机制SKNet,自适应地调整感受野大小,增强有用特征,抑制无用特征。最后采用激活函数Leaky-ReLU,减少沉默神经元的出现,避免了在输入为负值时使用ReLU导致的神经元失活。实验结果表明,算法在WTB数据集上的准确度达到了98.12%,比ShuffleNet V2提高了6.53%,且模型参数量仅为1.4 M。As the defect detection technology of wind turbine blades has not been widely used and the robust⁃ness of traditional detection methods is poor,this paper proposes a lightweight wind turbine blade surface defect i⁃dentification algorithm WTB-Net.2569 images of wind turbine blades from a coastal wind farm in East China are captured by an unmanned aircraft,and the WTB surface defect dataset is established through screening and classifi⁃cation and data expansion.Based on ShuffleNet V2 core backbone feature extraction network,SKNet,a selective convolutional kernel attention mechanism,is introduced to adaptively adjust the perceptual field size,enhance useful features and suppress useless features.Finally,the activation function Leaky-ReLU is used to reduce the ap⁃pearance of silent neurons and avoid the deactivation of neurons caused by using ReLU when the input is negative.The experimental results show that the algorithm achieves an accuracy of 98.12%on the WTB dataset,which is 6.53%better than ShuffleNet V2,and the number of model parameters is only 1.4 M.

关 键 词：深度学习 图像分类 ShuffleNet V2 风机叶片 复合材料 

分 类 号：TB332[一般工业技术—材料科学与工程]